Gemini 4 Argon 看一段视频就让 Apollo 机器人连中七投

Yeon Vail @yeonvail原推发布于 2026.10.15

在 X 上打开原推

google showed a model one video of a person playing cornhole. a humanoid then sank 7 throws in a row on real hardware. nobody hand-built the simulation it practiced in Gemini 4 Argon, announced by Google DeepMind on october 13 2026. it watched the clip, mapped the environment, worked out whole-body joint movements and wrote its own simulation code to practice. then the motion went to Apptronik's Apollo most will read this as "robot learns a party game". the part that matters is who built the practice room the idea: the model does not drive the robot. it builds the gym the robot trains in i went through the demos, the model card notes and the coverage under the purple lamp in praga what runs under the hood: 1. THE CORNHOLE: 1 human video → environment map → whole-body joint plan → self-written sim → Apollo sinks 7 straight. Google did not publish how many tries it took 2. THE SPOT: a Boston Dynamics Spot with 12-DoF legs and a 6-DoF arm. Argon builds a MuJoCo clone of a pedal bin, tests balance on 3 legs, then the real robot steps on the pedal, checks its wrist camera and drops the can in 3. THE PHOTO → GYM: 1 overhead photo of a table → calibrated sim twin → 77 tasks → 4,500+ randomized episodes. a robot policy trained on that data went from 56% to 75% real-world success and from 35 s to 24 s per task. Google does not name the robot 4. THE SKETCH: a paper sketch → a 76-DoF humanoid sim → a throwing trajectory → all 6 target boxes fall 5. THE SCORE: 87% on ERQA, the embodied reasoning benchmark. company-reported, not reproduced yet. the Live API is not available, so it plans and writes code while fast motor control stays with the robot's own stack 6. THE ACCESS: out now in the paid Gemini API as gemini-4-argon, in AI Studio and for Ultra subscribers. alone each demo is a party trick. together they say the slowest part of robot learning, building simulations by hand, just became a prompt. researchers say weeks of modeling → a few hours VIDEO / PHOTO / SKETCH → ARGON WRITES THE MUJOCO SIM → PRACTICE → POLICY → REAL ROBOT → CAMERA CHECK → RETRY my take: last month OpenAI Astra steered robot arms directly, and real-world testing on one benchmark stopped after it damaged hardware. Google went the other way. Argon never touches the motors, it builds the room where the motors practice. still, "7 in a row" has no denominator, Spot has no published success rate, and every number here is Google's own the novelty is not a robot throwing a beanbag. the novelty is that the most expensive person on a robotics team, the one who builds the simulator, just got a model that does it from an ordinary video bookmark this and watch who publishes the first independent 56 → 75 replication edge is open until every robot lab rebuilds its sim pipeline around one prompt

这条是引用着另一条写的：x.com/lctrnotes/status/2110434222091436412